bioRxiv preprint doi: https://doi.org/10.1101/871590; this version posted July 3, 2020. The copyright holder for this preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. TransINT: an interface-based prediction of membrane protein-protein interactions Khazen, G.1,*, Gyulkhandanian, A.2, Issa, T.1 and Maroun, R.C. 2,* 1 Computer Science and Mathematics Department, Lebanese American University, Byblos-Lebanon 2 Inserm U1204/ Université d’Evry/Université Paris-Saclay/ Structure-Activité des Biomolécules Normales et Pathologiques 91025 Evry France. ABSTRACT Membrane proteins account for about one-third of the proteomes of organisms and include structural proteins, channels, and receptors. The mutual interactions they establish in the membrane play crucial roles in organisms, as they are behind many processes such as cell signaling and protein function. Because of their number and diversity, these proteins and their macromolecular complexes, along with their physical interfaces represent potential pharmacological targets par excellence for a variety of diseases, with very important implications for the design and discovery of new drugs or peptides modulating or inhibiting the interaction. Yet, structural data coming from experiments is very scarce for this family of proteins and the study of those proteins at the molecular level goes necessarily through extraction from cell membranes, a step that reveals itself noxious/harmful for their structure and function. To overcome this problem, we propose a computational approach for the prediction of alpha-helical transmembrane protein higher-order structures through data mining, sequence analysis, motif search, extraction, identification and characterization of the amino acid residues at the interface of the complexes. This method leads us to the formulation of protein-protein binding sites used to scan protein sequence datasets for generating new potential interacting protein couples. 1 bioRxiv preprint doi: https://doi.org/10.1101/871590; this version posted July 3, 2020. The copyright holder for this preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. Our template motif-based approach using experimental recognition sites leads us to predict thousands of binary complexes between membrane proteins. We generate an online database of the annotated predicted interactions that can be accessed on https://transint.shinyapps.io/transint/. Key words: protein-protein interactions; bioinformatics; biostatistics; integral membrane proteins; computational biology; prediction method; binary interactions; molecular recognition; protein complexes 2 bioRxiv preprint doi: https://doi.org/10.1101/871590; this version posted July 3, 2020. The copyright holder for this preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. INTRODUCTION Proteins are the working horses of organisms and encompass a large variety of functions including, but not limited to, catalysis, transport, cell structure, and signal transduction. Our working definition of the membrane protein interactome is the complete set of direct interactions that take place among integral-to-the-membrane proteins (bitopic or single-spanning and polytopic or multiple-spanning). These transmembrane proteins (TM) span the cell membrane in its entirety and constitute the membranome, i.e. the proteome of the membrane. TMs represent around one third of the proteomes of organisms 1 and come mostly in the form of alpha-helix domains that cross different types of cell membranes. They include many important enzymes, receptors, and transporters (about 3 x 105 in number, excluding polymorphisms or rare mutations). The number of possible binary and multiple interactions between them is thus vastly larger. As proteins are the core of the cell machinery in organisms (activation, transport, degradation, stabilization, and participation in the production of other proteins), and their complexes are the functional units of cells, it is fundamental to understand the interactions between them. In other words, knowledge of the 3D structure of the proteins involved and of the interfaces needed for complex formation is a fundamental phenomenon governing all processes of life. 2 TM proteins represent potential pharmacological targets par excellence in human health and disease because they include many families involved in protein- protein interaction (PPI) networks, leading to many different physiological processes (signal transduction, apoptosis...). Thus, TM interactions, through the lipid-embedded domains, lead to oligomer formation and guide the assembly and function of many cell proteins and receptors. In addition, the assembly of TM proteins may lead to emergent properties, a relationship existing between oligomerization and function of these proteins. Indeed, deficient oligomerization is associated with known diseases. 3–5 Estimates of the total number of human PPIs range from 130,000 to 600,000 6–8 to 3,000,000 9, several orders of magnitude bigger than the D melanogaster interactome. High-throughput experimental and theoretical approaches are being used to build PPI networks and thus elucidate the rules that govern these interactions. However, traditional techniques like yeast two-hybrid (Y2H) assays 10 3 bioRxiv preprint doi: https://doi.org/10.1101/871590; this version posted July 3, 2020. The copyright holder for this preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. are not well suited for identifying membrane protein interactions. The data covering PPIs is increasing exponentially -in year 2012, there were more than 235,000 binary interactions reported. 11 Most protein interaction databases (DB) – bioGRID, 12 BIND, 13 STRING, 14 IntAct, 15 MINT, 11, and others - offer general information about experimentally validated PPIs of all types. The IMEx Consortium groups all nonredundant protein interaction data in one interface. 16 Nevertheless, these DBs are mostly concerned with water-soluble globular proteins and the juxtamembrane interactions of TM proteins. But, unlike globular proteins, TM proteins are water- insoluble and for technical reasons their large-scale investigation has lagged behind. 17 Proteome-wide maps of the human interactome network have been generated in the past. 18–20 Nevertheless, these assays, just like the Y2H assay, are depleted of membrane proteins or unreliable.21 And even though a new biochemical technique has been developed using a split-ubiquitin membrane two-hybrid (MYTH) system, so far only a very limited number of membrane complexes have been determined using it. 18,19 This procedure has been further extended with a mammalian-membrane two- hybrid assay (MaMTH) technique for the identification and characterization of the interaction partners of integral membrane proteins under desired conditions. 22 But to the best of our knowledge, MaMTH has not been used as a systematic screening assay to map the human membrane protein interactome. On the other hand, using a variety of classifiers there are many methods for the prediction of PPIs from: i) sequence alone (PIPE2; 23 SigProd; 24 MULTIPROSPECTOR; 25 PCA-EELM; 26 MCDPPI; 27 amino acid composition; 28; and ii) from template structures. For a review on the interactome, see. 29 But, again, most of the approaches address soluble globular proteins only.30 Qi et al developed a random forest learning machine approach but limited to the human membrane receptor interactome, HMRI. 31 On another hand, ab-initio prediction of membrane protein interfaces is rendered difficult as these have amino acid compositions that are not very different from the rest of the protein surface, decorated by hydrophobic residues in the membrane-exposed surface. To circumvent this problem, we developed a knowledge-based predictive approach based on the detection of alpha-helical transmembrane contact residues issued from membrane protein multimers, or their portions thereof reported in the PDB 32 and validated by the OPM DB, 33 in the context of the lipid bilayer. Querying 4 bioRxiv preprint doi: https://doi.org/10.1101/871590; this version posted July 3, 2020. The copyright holder for this preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. thereafter the PDBsum DB, 34,35 we obtain the atomic details of the protein-protein interfaces, i.e. the residues that are in contact at the interface of the complexes. We then gather those amino acids at the recognition interface to generate regular expressions, or patterns, that represent in a linear form the interaction motifs in space, including wildcard regions in addition to the interface residues. With this information in hand, we proceed to find the obtained motifs in other TM proteins by allowing a certain degree of amino acid mutations in the sequences of the motifs. The allowed mutation rates of the interface residues allow us to explore local, spatial non-homologous motifs. Homologs of the starting motifs are expected to interact in an analogous manner, 36 as opposed to a global structural approach that would find structurally, physically homologous PDB interfaces, limiting the search to functionally-related partners without paying attention to the particular sequence at the interface